Within the ever-expanding field of machine-learning (ML), new practical applications are found nearly everywhere. This study explores if ML-based demodulation can be used as an alternative to conventional frequency-modulation (FM) radio receivers, especially under non-ideal channel conditions. Specifically, a convolutional neural network architecture for regression is proposed to demodulate FM signals represented by their in-phase and quadrature components. A synthetic dataset of base-band FM signals, covering 20 Hz–20 kHz and augmented with a variety of typical signal disturbances, is used to train the model. Tests show that, when the incoming signal-to-noise ratio (SNR) is 10 dB or lower (the FM threshold region), the ML-model raises output SNR by up to 6 dB and reduces mean-squared error by an order of magnitude compared with the conventional method. Listening tests with 10 subjects corroborate these numerical gains. Because the model runs in real-time on a laptop CPU, ML-based demodulation could offer a practical route to more robust FM reception in noisy or adverse channels.
Karlsson et al. (Wed,) studied this question.